The AIUC-1 Compliance Checklist: 5 Layers Every Enterprise Needs Before Deploying AI Agents

A quick-reference checklist for AIUC-1 compliance. Five layers, 28 controls, one page. Print it, pin it, pass the audit. No single tool passes AIUC-1. You need five layers. Here’s the checklist. Layer 1: Control Plane (Access & Enforcement Gateway) Deploy first. Everything else depends on this. Centralized gateway deployed — all agent-to-tool traffic flows through a single enforcement point Tool-level authorization active — agents can only access specific tools they’re authorized for , not entire servers Identity-aware policies configured — every request is authenticated via SSO/JWT with verified user context Session tracking enabled — multi-step agent workflows tracked as continuous sessions Comprehensive audit logging active — every request logged with : agent identity, user context, tool called, parameters, policy decision, timestamp, session ID Rate limiting configured — per-agent and per-route request limits prevent scraping and DoS Multi-tenant isolation validated — Customer A’s agents cannot access Customer B’s data Deployment model assessed — cloud vs. self-hosted decision documented AIUC-1 controls covered: B006, B004, B007, A003, A005, E004, E015, E005 Tool: Pomerium Agentic Gateway Layer 2: Model Safety & Content Filtering Controls what agents output — independent of what they access. Input filtering deployed — prompt injection and jailbreak attempts caught before reaching the model Output filtering deployed — harmful, toxic, biased, or deceptive content blocked before reaching users PII detection active — personal data in agent outputs is detected and redacted Hallucination detection enabled — fabricated facts and data flagged before delivery CBRN guardrails configured — content policies block chemical, biological, radiological, and nuclear misuse Filtering events logged — every blocked or flagged input/output is logged with reason AIUC-1 controls covered: C003, D001, B005, F001, F002 Tools: Azure AI Content Safety, AWS Bedrock Guardrails, Protect AI, Anthropic/OpenAI built-in safety Layer 3: Testing & Red-Teaming Quarterly validation that Layers 1 and 2 actually work. Adversarial testing program established — quarterly red-teaming with documented methodology Pre-deployment testing required — all agent changes tested against risk categories before production Third-party safety evaluation scheduled — external firm evaluates safety at least quarterly Hallucination benchmarks tracked — hallucination rates measured and compared to baseline quarterly Regression suite automated — previous adversarial findings re-tested every quarter Test results documented — pass/fail, findings, remediation steps, sign-off AIUC-1 controls covered: B001, C002, C004, D002 Services: Schellman (authorized AIUC-1 auditor), HackerOne AI, Bishop Fox, internal security teams Layer 4: Observability & Detection Turns raw logs into intelligence and compliance reports. Log aggregation pipeline configured — gateway logs, model logs, and application logs centralized Anomaly detection rules active — unusual request patterns, suspected attacks flagged automatically Compliance dashboards built — real-time view of: top agents by request count, denied requests, data access trends Alerting configured — automated alerts for policy violations, rate limit breaches, suspected exfiltration Audit report generation automated — quarterly compliance reports generated from aggregated data AIUC-1 controls covered: B002, E015 (augmented) Tools: Splunk, Elastic, Microsoft Sentinel, Datadog, Grafana Layer 5: Governance & Process The organizational layer auditors evaluate alongside technology. Data use policies written — what agents can access, how outputs can be used, retention periods Risk taxonomy defined — formal classification of AI risks (hallucination, misuse, PII leak, adversarial attack) Incident response plans created — playbooks for: data breach, harmful output, hallucination in sensitive context Accountability matrix documented — RACI for every AIUC-1 control with named owners Vendor due diligence process established — evaluation checklist for model providers, tool vendors, auditors CBRN safeguards documented — written guardrails against catastrophic misuse scenarios All governance docs reviewed and signed off — CISO, legal, business leadership approval AIUC-1 controls covered: A001, A002, C001, E001, E002, E003, E004, E006, F001, F002 Implementation Timeline Phase Weeks Focus Controls Phase 1 1–4 Deploy control plane + audit logging B006, B004, B007, A003, A005, E004, E015, E005 Phase 2 5–8 Content filtering + model safety C003, D001, B005, F001/F002 Phase 3 9–12 Governance documentation A001, A002, C001, E001–E003, E006 Phase 4 Ongoing Quarterly testing cycle B001, C002, C004, D002 Quick Self-Assessment Answer these five questions. If any answer is “no,” you have a gap: Does every agent request flow through a centralized enforcement point? (If no → you need Layer 1) Are agent outputs filtered for harmful content, PII, and hallucinations? (If no → you need Layer 2) Do you run adversarial tests at least quarterly? (If no → you need Layer 3) Can you generate a compliance report from your logs in under an hour? (If no → you need Layer 4) Does every AIUC-1 control have a documented owner? (If no → you need Layer 5) Start With Layer 1 The most common mistake is starting with governance (Layer 5) or observability (Layer 4) before deploying a control plane (Layer 1). Without centralized enforcement and logging, everything else is unverifiable. Deploy the control plane first. The rest follows. Get started with Pomerium → Further reading: NIST AI Risk Management Framework · MITRE ATLAS · OWASP LLM Top 10 · AIUC-1 Standard